Willing but Unable: Separating Refusal from Capability in Code LLMs via Abliteration

1. 1Build a dataset of labeled vulnerable code using existing methods
2. 2Apply instruction-tuning to an LLM to inject specified vulnerabilities
3. 3Configure ablation to separate refusal from capability in code LLMs
4. 4Test the effectiveness of the LLM in detecting vulnerabilities
5. 5Run experiments to evaluate the impact of ablation on vulnerability detection

Security engineers and researchers on a team benefit from understanding how to evaluate code LLMs' capabilities and limitations, and how to improve vulnerability detection

💡 Ablation can help distinguish between an LLM's refusal to perform a task and its inability to do so, improving vulnerability detection

🚨 Improve code security by separating refusal from capability in LLMs via ablation 💻